EA How To
Build an AI Agent by Jeremy Dalton

The Assistant’s Assistant

Four prompts.
One working agent.

Work down the page in order. Edit the details in Step 1, copy each prompt, and paste it into Claude when instructed.

Start with Step 1 Open Claude

Before you begin

Keep these two windows open

1

Add the Project Instructions

In your Project, select Instructions. Complete the fields in MY SETUP, use none for optional items you do not use, and leave all other prompt text unchanged.

Personalise this one. At minimum, complete EMAIL, TIMEZONE, MY EMAIL and URGENT WINDOW.

Setup builder

Choose your systems

Open the builder to personalise the prompt.

Each choice updates the Project Instructions below. You can still edit the prompt directly for names, executives, VIPs or a custom value.

Only select optional systems that are already connected to Claude. Selecting one here does not create the connection.

Required

Optional

Choose your email system and enter your email address to enable Copy prompt.

Project Instructions
2

Create the Log and Test Connections

Start a Cowork conversation inside the Project and paste this Core Stage Gate. It creates the log if needed, then tests the log, email and every configured optional connector. Optional failures are reported without blocking the core build.

Verify the setup
Your finish line CORE ALL CLEAR: READY TO LOAD INITIAL CONTEXT

Fix a failed optional connector or set it to none before the Context Scan.

3

Run the Initial Context Scan

Paste this into the same Cowork conversation after the Stage Gate. Run it once to build the starting record. Do not schedule this prompt.

Load recent context
Your finish line INITIAL CONTEXT SCAN: COMPLETED

INITIAL CONTEXT SCAN: PARTIAL is also safe when Claude clearly identifies a source failure or backlog.

4

Create the Daily Start Brief

Open Scheduled → New task → Set up manually. Name it Daily Start Brief, choose a suitable schedule, keep permissions on manual approval, and paste this as its instructions.

Schedule Weekdays at 09:00
Model Standard
Permissions Manually approve
Daily Start Brief instructions

Now run it twice

  1. Run one: check for a CONFIG: 1.3 | RUN MODE | LOG MODE | TAG STYLE line matching the resolved setup.
  2. Run two: check that no labels, commitments or goals were duplicated.

The human loop

Confirm a few goals

After the second run, reply in that session. Change the numbers to match the result you received.

Example reply

Core workshop complete

You have built one agent with one proactive routine.

Additional routines and automatic approval come after the core double-run is clean.

Return to Claude

Optional routines

Keep building the same agent

Add these inside the same Project after the Daily Start Brief passes its double-run test. Create each as a separate Scheduled Task, begin with manual approval, and run it twice before allowing automatic approval.

Optional routine

Advisory Digest

Reviews upcoming calendar risks, response chasers, RSVP chasers, meeting preparation and due nudges.

Schedule Weekdays at 13:00
Model Standard
Permissions Manually approve first
Advisory Digest instructions

Optional routine

End of Day Review

Checks progress against today's goals and carries unresolved work forward without guilt or duplication.

Schedule Weekdays at 17:00
Model Standard
Permissions Manually approve first
End of Day Review instructions

Optional routine

End of Week Wrap

Summarises wins, open commitments and risks, then prepares a practical launchpad for Monday.

Schedule Friday at 16:30
Model Top model if available
Permissions Manually approve first
End of Week Wrap instructions
Jeremy Dalton

Created by

Jeremy Dalton

AI & Emerging Tech
Advisor, Author and Speaker

jeremy@xrhq.com www.jeremydalton.ai